Production routeThis rate reflects platform-side availability — only confirmed server errors (HTTP 500 / empty response) count as failures. User-side issues (content moderation, invalid params, cancellation) plus rate limits, timeouts and auth errors are excluded. Before real traffic arrives, empty buckets may display as available.Live
Live
Context
1M context · 384K max output
Best For
Screenshot and document understanding, visual QA, multimodal agents
Input
Text + images
Output
Text · JSON (structured output)
On EvoLink, this ID now redirects to DeepSeek V4.1 Flash; use deepseek-v4.1-flash for new integrations. Read the migration guide before reusing earlier image results.
A native multimodal route that reads text and images in the same request, available on Chat Completions, Messages, and Responses — use whichever protocol your stack already speaks.
DeepSeek V4 Flash Vision
Legacy Vision Exp ID, now routed to DeepSeek V4.1 Flash
Selected
Model ID
deepseek-v4-flash-vision-exp
Best for
Screenshot and document understanding, chart and table extraction, OCR and UI pipelines, plus any text workload that occasionally needs to read an image.
Input
$0.265 / 1M-10%
18 cr / 1M$0.295official price
Cache hit
$0.0059 / 1M
0.4 cr / 1M
Output
$1.059 / 1M-10%
72 cr / 1M$1.177official price
All rates are per 1M tokens, shown in USD and credits, and reflect your account's current pricing.
DeepSeek V4 Flash Vision pricing
Estimate what one DeepSeek V4 Flash Vision request costs before you integrate. The calculator uses your account's current rates, with official pricing as a reference.
On EvoLink, requests to deepseek-v4-flash-vision-exp now redirect to DeepSeek V4.1 Flash, after DeepSeek retired the experimental Vision Exp model on September 10, 2026. Existing image requests keep working on the same ID. For new integrations, use deepseek-v4.1-flash, and re-run your image evaluation set before relying on earlier results.
DeepSeek V4 Flash Vision is served on EvoLink under the model ID deepseek-v4-flash-vision-exp through Chat Completions · Anthropic Messages · Responses, with the same API key and balance you use for every other model. It offers a 1M context window and up to 384K output tokens, plus image input, long-context coding, thinking mode.
DeepSeek V4 Flash Vision specs and capabilities
Numbers come from the EvoLink route configuration; capabilities are what the API exposes today.
Context window
1M tokens
Max output
384K tokens
Input
Text + images
Output
Text · JSON (structured output)
Reasoning
Optional thinking mode
Prompt caching
Automatic cache reads at a lower rate
Protocols
Chat Completions · Anthropic Messages · Responses
Model ID
deepseek-v4-flash-vision-exp
Image input
PNG · JPEG · WebP · GIF
Requests on Chat Completions, Messages, and Responses must use the exact model ID shown here.
What is the DeepSeek V4 Flash Vision API best suited for?
Image understanding on this ID is now served by DeepSeek V4.1 Flash. Start with visual tasks that have clear acceptance criteria, then compare quality, latency, retries, and cost per accepted result before scaling.
Screenshot and UI inspection
Send product screenshots, dashboards, or error states with a focused question. Extract visible labels, identify layout or state differences, and return a structured QA summary. Validate small text, coordinates, and dense interfaces against a fixed screenshot set before automating decisions.
Document and receipt extraction
Turn scanned forms, invoices, receipts, and report pages into structured text for review or downstream processing. Define the expected JSON fields, preserve the source image for audit, and route low-confidence or missing fields to human review.
Chart and diagram analysis
Ask questions about charts, architecture diagrams, and technical figures while keeping the visual evidence in the same request. Require the response to distinguish observed labels from interpretation, especially when values are small or visually ambiguous.
Visual agents and UI verification
Combine screenshots with tool results so an agent can inspect a page state, propose the next action, or verify whether a workflow reached the expected screen. Keep actions reversible and evaluate task completion, latency, retries, and cost before expanding traffic.
What changes when you access DeepSeek V4 Flash Vision through different platforms?
Even with the same model, platforms can differ in context configuration, tool support, usage reporting, and billing. EvoLink puts model configuration and usage behind one API so later model changes remain straightforward.
Use the correct API model ID
The page URL and the API request model ID are the same string: deepseek-v4-flash-vision-exp. Existing Chat Completions or Responses applications can use the correct ID in model configuration, with exact fields available in the API section.
Follow the active API route configuration
DeepSeek documents a 1M-token model window, but platforms may expose different context settings, tools, and rate limits. When using EvoLink, rely on the model configuration, available features, and actual usage shown for the current route.
Choose an image-input protocol by workflow
DeepSeek documents image inputs for Chat Completions, Messages, and Responses, but gateway exposure can differ. Use only the protocols and content-block formats currently listed for this model in EvoLink documentation, then test streaming, usage, and error behavior before rollout.
Keep model choice open in one gateway
Use one EvoLink account, balance, and API pattern for Grok, GPT, Claude, and Kimi. Keeping model selection in configuration lets teams route by task quality, cost, and availability without rebuilding application code for every provider.
How can you control DeepSeek V4 Flash Vision API cost more accurately?
The existing Pricing section shows current token rates without duplicating a second price table. For the original Vision Exp model DeepSeek documented up to 384 input tokens per image; its current Vision guide sets an upper bound of 1024 tokens per image on the direct API, and requests to this ID now run on V4.1 Flash. Combine the current rule with live EvoLink rates, output, retries, and cache behavior when estimating completed-task cost.
Budget image input before output
Use DeepSeek’s current upper bound of 1024 tokens per image on its direct API as a planning ceiling, then inspect actual usage on representative screenshots and documents. Multiple images, long prompts, output, and retries still contribute to the final charge.
Make repeated context cache-friendly
Stable system prompts, tool schemas, and shared context are easier to reuse through caching. Configure the supported cache or conversation identifier for the selected protocol and inspect cached tokens in usage to confirm that repeated requests receive the expected benefit.
Send only the context the task needs
A 1M-token window is useful for large repositories and documents, but it does not need to be filled on every request. Selecting only relevant files, messages, and retrieved passages reduces input cost and helps the model focus on the evidence that matters.
Compare models by total task cost
Evaluate tokens, cached input, server tools, and required retries within the same completed task. EvoLink centralizes model and usage information so teams can compare the total cost of completing equivalent work across DeepSeek V4 Flash Vision and other routes.
Two ways to use DeepSeek V4 Flash Vision: EvoLink API or Agent
Use the EvoLink API for product backends and batch jobs, or call DeepSeek V4 Flash Vision from Codex, Claude, or Gemini for coding and analysis workflows. Both paths share the same EvoLink API key, balance, model ID, and request history.
Option 1
Integrate with the EvoLink API
Best for: product backends, batch jobs, automated pipelines
Send OpenAI-compatible Chat Completions (or Anthropic Messages) requests to EvoLink and control the model ID, system prompt, output budget, tools, and structured output.
1Create an EvoLink API key in the console
2Point your OpenAI or Anthropic SDK at the EvoLink base URL and select the model ID shown above
3Send one representative request and read the usage field for input, cached, and output tokens
4Set max_tokens and retries per task; keep tool-call IDs and results across turns
Best for: coding, review, and analysis tasks in Codex, Claude, and Gemini
Give the Agent the task, the inputs to include, and the acceptance criteria. It assembles the request, calls DeepSeek V4 Flash Vision through EvoLink, and returns the answer with token usage.
1Set EVOLINK_API_KEY in your local environment; never put it in code or a prompt
2Describe the task, the inputs to include, and the expected output format
3Ask the Agent to call DeepSeek V4 Flash Vision through EvoLink and show the request before sending
4Let the Agent report the answer, token usage, and any error body
DeepSeek V4 Flash Vision API code example and error handling
This example shows the shortest runnable request: an OpenAI-compatible Chat Completions call with a system prompt, a user message, and an output budget. Open the API tab for the complete parameter and response reference.
curl -X POST https://api.evolink.ai/v1/chat/completions \
-H "Authorization: Bearer $EVOLINK_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "deepseek-v4-flash-vision-exp",
"messages": [
{ "role": "system", "content": "You are a senior engineer reviewing a pull request." },
{ "role": "user", "content": "Review this diff and list every risky change with file and line:\n<diff>" }
],
"max_tokens": 4096,
"temperature": 0.2
}'
# Anthropic Messages (/v1/messages) and Responses (/v1/responses) accept the
# same model ID.
# The response includes choices[0].message and a usage object (prompt,
# cached, and completion tokens).
Invalid request or unsupported parameter
Check the model ID, the messages array, and parameter ranges against the API reference; remove fields this route does not support.
Authentication or balance issue
Check the Authorization bearer token and confirm the available balance in the console.
Context length exceeded
Prompt tokens exceed the DeepSeek V4 Flash Vision context window. Trim or retrieve only the relevant evidence and reuse cached prefixes.
Rate limited (429)
Back off and retry with jitter; batch or queue requests instead of sending parallel bursts.
Content or tool call rejected
Review sensitive content, malformed tool-call arguments, and JSON schema mismatches before retrying.
How should you choose an image-input protocol and production route for DeepSeek V4 Flash Vision?
DeepSeek documents three image-capable protocol shapes. This section explains the decision; the current EvoLink API documentation remains the authority for which model IDs, endpoints, and content blocks are enabled on the gateway.
01
Chat Completions: familiar image_url content
Use Chat Completions when your application already follows the OpenAI-style messages array. Confirm that the Vision Exp model and image_url content type appear in the current EvoLink docs before using the request shape in production.
OpenAI-style chat
02
Messages: Anthropic-style image blocks
Use Messages only after EvoLink documentation explicitly lists image content blocks for deepseek-v4-flash-vision-exp. Include the documented max_tokens field and verify how usage, thinking content, and errors are returned.
Anthropic-style messages
03
Responses: multimodal agent input
Evaluate Responses when a longer agent workflow benefits from input_image and structured response handling. Confirm the active request fields, supported tools, streaming events, and retry boundary against EvoLink documentation.
Agent workflows
04
Production traffic: start small and keep a fallback
Send a small, observable set of tasks to DeepSeek V4 Flash Vision first and retain a proven GPT, Claude, or Kimi route. EvoLink’s unified API centralizes model choice, usage, and balance, making it easier to change routes after rate limits, timeouts, schema errors, or tool failures.
Gradual rollout
Cache hits are billed at a fraction of fresh input
Prefix caching is automatic: a stable system prompt, repository instructions, and tool definitions are billed as cache hits on later requests instead of full input. Measure the real hit rate in your logs before estimating agent-loop cost.
1M-token context with a 384K output ceiling
Keep related files, specifications, and agent state in one working context, but treat the ceiling as capacity rather than a target: retrieve what matters, keep the prefix stable so it caches, and set an output budget per task.
Compare DeepSeek V4 Flash Vision with Qwen 3.8 Max and GLM-5.3
EvoLink
Compare input/output rates, context, caching, and workload fit. Benchmark the same requests before choosing a route; prices below follow your account’s current rates.
Screenshot and document understanding, chart and table extraction, OCR and UI pipelines, plus any text workload that occasionally needs to read an image.
Coding agents, long task chains, tool-heavy workflows
DeepSeek V4 Flash Vision
Input / output$0.265 / $1.059
Context1M
CachingAutomatic cache hits
Best forScreenshot and document understanding, chart and table extraction, OCR and UI pipelines, plus any text workload that occasionally needs to read an image.
Best forCoding agents, long task chains, tool-heavy workflows
What should you confirm before using DeepSeek V4 Flash Vision in production?
Start with a small set of real workloads to determine whether the model meets your quality, latency, cost, and reliability needs before deciding which traffic should move.
Integration and usage data are clear
Confirm that the application uses the correct deepseek-v4-flash-vision-exp model ID and intended protocol, and that required responses, usage, and cache information are returned. Clear usage data supports cost analysis and gives Chat and Responses workflows a consistent observation method.
Visual outputs meet real business requirements
Test representative screenshots, forms, charts, and UI states. Check field accuracy, missing values, small-text handling, hallucinated details, structured-output validity, and the amount of human correction required before accepting a result.
Latency and error handling meet expectations
Observe response time under representative traffic and prepare retry behavior for rate limits, timeouts, invalid structured output, and tool failures. Configurable model selection in EvoLink makes it easier to switch to a verified alternative when one route is temporarily unavailable.
Cost and model choice remain controllable
Use Pricing, usage, and final charges to calculate total cost for the same class of task, then decide whether DeepSeek V4 Flash Vision belongs on the default route, difficult tasks, or fallback traffic. A unified gateway keeps quality and budget decisions separate from integration work.
Begin with a small, observable, reversible set of DeepSeek V4 Flash Vision tasks. Expand only after quality, latency, and cost meet expectations. Keeping multiple model options behind the EvoLink unified API makes later scaling, switching, and cost optimization easier.
DeepSeek V4 Model Family
Same API key and balance — switch tiers without changing your integration.
DeepSeek V4.1 Flash
DeepSeek’s multimodal Flash model for text and image workloads
The text-only V4 Flash model on EvoLink, without image input. Check its own Pricing section for current rates, and use it when your workload never sends images.
What is the DeepSeek V4 Flash Vision Exp model ID?
The original model ID is deepseek-v4-flash-vision-exp, released as an experimental model on August 21, 2026. On EvoLink the ID still works but now redirects to DeepSeek V4.1 Flash; new integrations should use deepseek-v4.1-flash.
Is image input available through EvoLink?
Use the current EvoLink API documentation as the route authority. The upstream DeepSeek model accepts image input, but each EvoLink protocol and content-block format should be considered enabled only when it is listed in the current docs and passes a real request on your account.
Which image formats does the upstream model support?
DeepSeek documents JPEG, PNG, GIF, and WebP input. Check EvoLink documentation for gateway-specific URL, Base64, file-size, and multi-image requirements before production use.
How many input tokens does one image use?
DeepSeek documented up to 384 input tokens per image for the original Vision Exp model. Its current Vision guide sets an upper bound of 1024 tokens per image on the direct API, and requests to this ID now run on V4.1 Flash. Plan with the current rule, then inspect actual usage and the live EvoLink Pricing section on representative images; prompts, multiple images, output, retries, and cache behavior also affect the final charge.
Should I use Chat Completions, Messages, or Responses?
Choose the protocol your application already uses, provided EvoLink currently documents image support for this model on that protocol. Chat uses image_url-style content, Messages uses image blocks, and Responses uses input_image upstream; verify the active EvoLink request shape before implementation.
When should I use Vision Exp instead of V4 Flash?
For new image work on EvoLink, use deepseek-v4.1-flash; requests to deepseek-v4-flash-vision-exp already redirect to it. Keep deepseek-v4-flash for text-only, throughput-oriented workloads: it is not affected and still serves V4 Flash.
Can I rely on it for dense screenshots or small text?
Treat dense UI, small text, coordinates, and crowded tables as evaluation cases, not guaranteed strengths. Build a fixed test set, compare extracted fields against ground truth, and route low-confidence results to review.
What are the rate and file-size limits?
Do not infer limits from the text-only Flash route or the upstream API. Check the current EvoLink model documentation and dashboard for route-specific limits, then test 429, timeout, and oversized-image behavior before scaling.
How should teams compare Vision Exp with other vision models?
Run the same images, prompts, output schema, and review criteria across routes. Compare field accuracy, hallucinations, latency, retries, and total cost per accepted result rather than relying on a vendor benchmark claim.
What fallback should teams keep during rollout?
Keep a verified model that handles the same visual workload and leave model selection configurable. Text-only models such as V4 Flash cannot take over image tasks. Start with a small traffic share, monitor errors and acceptance rate, and switch routes when the current model misses your production threshold.
